Abstract: To address the instability and non-real-time problems of manual quality control in cerebral angiography and realize real-time quality assessment, a lightweight segmentation-classification multi-task learning (MTL) model is proposed. This model comprises three main components: a feature extraction backbone module, a blood vessel segmentation module, and an angiography quality classification module. Depth-separable convolution is employed instead of traditional convolution to reduce the number of parameters. Additionally, a local-global self-attention module (L-GSAM) is proposed to enhance the model’s ability to extract global information. A feature aggregation module (FAM) is introduced in the vessel segmentation module to optimize feature connections. The segmentation results are then combined with backbone features to assess the quality of angiography in the classification module, and a joint loss function is designed for model training. Experimental results show that the proposed model achieves good segmentation and classification performance with only 3.4342×10⁶ parameters, the quality assessment accuracy reaches 0.818 2, and the model exhibits high real-time performance.

Key words: cerebral angiography, multi-task learning, local-global self-attention, feature aggregation